Consensus Focus for Object Detection and minority classes

Ensemble methods exploit the availability of a given number of classifiers ordetectors trained in single or multiple source domains and tasks to addressmachine learning problems such as domain adaptation or multi-source transferlearning. Existing research measures the domain distance between the sourcesand the target dataset, trains multiple networks on the same data withdifferent samples per class, or combines predictions from models trained undervaried hyperparameters and settings. Their solutions enhanced the performanceon small or tail categories but hurt the rest. To this end, we propose amodified consensus focus for semi-supervised and long-tailed object detection.We introduce a voting system based on source confidence that spots thecontribution of each model in a consensus, lets the user choose the relevanceof each class in the target label space so that it relaxes minority boundingboxes suppression, and combines multiple models' results without discarding thepoisonous networks. Our tests on synthetic driving datasets retrieved higherconfidence and more accurate bounding boxes than the NMS, soft-NMS, and WBF.